Hi Mark,
Thanks for the reply.
Mark Jenkinson wrote:
> Hi,
>
> If you are dealing with images that are almost aligned at the
> beginning then I find
> it quite worrying that they do not register well. Are you using the
> -nosearch option?
> This turns off the initial (large) search phase, but allows the
> perturbations and
> local refinement of the transformation.
Usually I do not use the -nosearch option, but I set the -searchrx\y\z
quiet low (~ 3 deg) to account for small differences between sessions. I
only use the -nosearch option if I alligned the data already to a common
template.
>
> Also, are there significant distortions between the images you are
> registering?
The last time when I encounter this problem I was coregistering a
magniude image of the fieldmap (low resolution) with a FLASH image (high
resolution), using a bet masks of both modalities as weight. So, here
are not much distortions in the data, but maybe due to the resolution
difference flirt thought that the magnitude bet image corresponds just
to the cerebellum. I improved the results by not using the weight mask.
I also reduced DOF to 6 again.
> It is possible that it cannot find a good affine registration because
> of this.
> Are you using more than 7 dof because of these distortions? If so, a
> fieldmap
> approach (unwarping the images using the fieldmap information) would be
> preferable if you have them.
In the EPI images there are of course nice distortions since it is awake
behaving monkey data. I tried epi undistortions a while ago, but was not
very happy with the result. In general I got improvements of the global
brain shape, but locally I introduced severe artefacts. I interpreted it
as a problem resulting from dynamically changes in the field due to
swallowing (large muscle movements) and body movements. As long the
fieldmap is not acquired close in time to the epi I would expect
problems. However, if I got some spare time I want to try out a few more
things. However, I had the impression that the 12 DOF affine
transformation gave very good results with the coregistrations.
Functional maps are usually pretty well located in the grey matter and
allow for a reliable localization. Fortunately the problem I described
here does not occur very often (< 10 %) and I can cope with it by
manually adjusting the coregistration parameter.
If you want to try something with the data I could select some files and
upload them, but I am not sure how fast I could manage it, because right
now I should have some holidays. Anyway, I'll prepare a selection soon.
Thanks,
wolf
>
> There is no explicit constraint on the parameters. If you cannot get
> a good
> solution at present this is really unlikely to help as it will most
> likely just give
> you a solution where it is against the hard limit on one or more
> parameters
> which may not be any better than your original position. You can
> explicitly
> set the scaling by passing in an initial matrix where the scaling is
> set (via
> the -init option) and setting the dof to 6 so that it will not change
> the scaling.
> However, for the reasons above I do not really recommend this. Certainly
> I would not recommend trying things with schedule files.
>
> If your images are not significantly distorted and are of the same
> individual
> and you still cannot get a good registration with 6 (or 7) dof and the
> -nosearch
> option (or, better, using fieldmap unwarping) then feel free to send
> us some
> example images for us to try via our upload site:
> http://www.fmrib.ox.ac.uk/cgi-bin/upload.cgi
>
> All the best,
> Mark
>
>
> On 26 Dec 2008, at 00:11, wolf zinke wrote:
>
>> Hi,
>>
>> Here is something for my Xmas wishlist - I noticed that flirt
>> sometimes gives some weird results when using a 9 or 12 DOF
>> registration. It manages to put the whole brain for example into the
>> cerebellum of the reference image. I guess this misbehaviour would be
>> improved if there are restrictive constraints that limit the scaling.
>> Another option allowing to specify the scaling explicitly might be
>> very helpful (as it is realized for the rotation angles). It also
>> could help if this is done for all other fitting parameter as well,
>> especially for the translation. Since I am working with primates
>> having their head fixed I do not expect much real movement besides
>> the artefacts caused by body displacements. The latter result
>> primarily in an apparent head displacement in phase encode direction.
>> Therefore I assume that using explicit constraints for each parameter
>> for flirt and mcflirt might give even better results. Maybe this
>> finer control is already available with the schedule files. In this
>> case I am sorry that I didn't spent much time yet to understand the
>> structure of these schedule files and to do this would then be one of
>> my New Year's pledges.
>>
>> Thanks a lot for all the great tools,
>> wolf
|